[
https://issues.apache.org/jira/browse/YARN-1773?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Bikas Saha updated YARN-1773:
-----------------------------
Description:
Currently, the ShuffleHeader (which is a Writable) simply tries to read the
successful header (mapid, reduceid etc). If there is an error then the input
will have an error message instead of (mapid, reducedid etc). Thus parsing the
ShuffleHeader fails and since we dont know where the error message ends, we
cannot consume the remaining input stream which may have good data from the
remaining map outputs. Being able to encode the error in the ShuffleHeader will
let us parse out the error correctly and move on to the remaining data.
The shuffle handler response should say which maps are in error and which are
fine, what the error was for the erroneous maps. These will help report
diagnostics for easier upstream reporting.
was:Currently, the ShuffleHeader (which is a Writable) simply tries to read
the successful header (mapid, reduceid etc). If there is an error then the
input will have an error message instead of (mapid, reducedid etc). Thus
parsing the ShuffleHeader fails and since we dont know where the error message
ends, we cannot consume the remaining input stream which may have good data
from the remaining map outputs. Being able to encode the error in the
ShuffleHeader will let us parse out the error correctly and move on to the
remaining data.
> ShuffleHeader should have a format that can inform about errors
> ---------------------------------------------------------------
>
> Key: YARN-1773
> URL: https://issues.apache.org/jira/browse/YARN-1773
> Project: Hadoop YARN
> Issue Type: Bug
> Affects Versions: 2.3.0
> Reporter: Bikas Saha
> Priority: Critical
>
> Currently, the ShuffleHeader (which is a Writable) simply tries to read the
> successful header (mapid, reduceid etc). If there is an error then the input
> will have an error message instead of (mapid, reducedid etc). Thus parsing
> the ShuffleHeader fails and since we dont know where the error message ends,
> we cannot consume the remaining input stream which may have good data from
> the remaining map outputs. Being able to encode the error in the
> ShuffleHeader will let us parse out the error correctly and move on to the
> remaining data.
> The shuffle handler response should say which maps are in error and which are
> fine, what the error was for the erroneous maps. These will help report
> diagnostics for easier upstream reporting.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
